AI Paper Summary

Contrastive Language-Image Pre-training (CLIP) has become important for modern vision and multimodal models, enabling applications such as zero-shot image classification and serving as vision encoders in MLLMs. However, most CLIP variants, including Meta CLIP, are limited to English-only data curation, ignoring a significant amount of non-English content from the worldwide web. Scaling CLIP to include…

Reinforcement learning (RL) plays a crucial role in scaling language models, enabling them to solve complex tasks such as competition-level mathematics and programming through deeper reasoning. However, achieving stable and reliable training dynamics is a challenge when scaling RL with larger computational resources. Current state-of-the-art algorithms, such as GRPO, struggle with serious stability issues during…

LLMs are deployed through conversational interfaces that present helpful, harmless, and honest assistant personas. However, they fail to maintain consistent personality traits throughout the training and deployment phases. LLMs show dramatic and unpredictable persona shifts when exposed to different prompting strategies or contextual inputs. The training process can also cause unintended personality shifts, as seen…

LLMs have shown notable improvements in mathematical reasoning by extending through natural language, resulting in performance gains on benchmarks such as MATH and AIME. However, reinforcement learning (RL) for training these models encounters a challenge: verifying the correctness of natural language proofs is very difficult, requiring careful manual checking of each reasoning step. This limits…